Chapter 4:

Fuzzy Inference Systems

 Introduction (4.1)
 Mamdani Fuzzy models (4.2)
 Sugeno Fuzzy Models (4.3)
 Tsukamoto Fuzzy models (4.4)
 Other Considerations (4.5)
–Fuzzy modeling
 Introduction (4.1) 2

 Fuzzy inference is a computer paradigm

based on fuzzy set theory, fuzzy if-then-
rules and fuzzy reasoning
Applications: data classification, decision
analysis, expert systems, times series
predictions, robotics & pattern recognition
Different names; fuzzy rule-based system,
fuzzy model, fuzzy associative memory, fuzzy
logic controller & fuzzy system
Fuzzy System: (Rule Base, Data Base,
Reasoning Mechanism)
 Introduction (4.1) 3

 Structure
– Rule base  selects the set of fuzzy rules
– Database (or dictionary)  defines the
membership functions used in the fuzzy rules
– A reasoning mechanism  performs the
inference procedure (derive a conclusion from
facts & rules!)
 Defuzzification: extraction of a crisp value that
best represents a fuzzy set
– Need: it is necessary to have a crisp output in
some situations where an inference system is
used as a controller
 4
Rule 1
 w1
x is A1 y is B1
Rule 2
 defuzzifier
x is A2 y is B2
 w2
x aggregator y
Rule r
 is A y is Br

x x r Fuzzy
wr

Block diagram for a fuzzy inference system

Non linearity 5
In the case of crisp inputs & outputs, a fuzzy
inference system implements a nonlinear
mapping from its input space to output space
Mamdani Fuzzy models [1975](4.2)
 Goal: Control a steam engine & boiler
combination by a set of linguistic control
rules obtained from experienced human
operators
Illustrations of how a two-rule Mamdani
fuzzy inference system derives the overall
output z when subjected to two crisp input
x&y
 6

The Mandamy fuzzy inference using min and

max for T-norm and T-conorm operators,
respectively
 Mamdani Fuzzy models (4.2) (cont.) 7

Defuzzification [definition]

“It refers to the way a crisp value is extracted from

a fuzzy set as a representative value”

– There are five methods of defuzzifying a fuzzy

set A of a universe of discourse Z

• Centroid of area zCOA

• Bisector of area zBOA
• Mean of maximum zMOM
• Smallest of maximum zSOM
• Largest of maximum zLOM
Mamdani Fuzzy models (4.2) (cont.) 8
• Centroid of area zCOA   A ( z )zdz
z COA  Z
,
  A ( z )dz
Z
where A(z) is the aggregated output MF.
• Bisector of area zBOA
this operator satisfies the following;
z BOA 

  A ( z )dz    A ( z )dz ,
 z BOA
where  = min {z; z Z} &  = max {z; z Z}.
The vertical line z = zBOA partitions the region
between
z = , z = , y = 0 & y = A(z) into two regions
 Mamdani Fuzzy models (4.2) (cont.) 9
Mean of maximum zMOM
This operator computes the average of the
maximizing z at 
which the MF reaches a maximum  . It is
expressed by :
 zdz
z MOM  Z'
,
 dz
Z'

where Z'  { z;  A ( z )   }

By definition : if  A ( z ) has a single maximum at z  z

then z MOM  z
z1  z 2
However : if max  A ( z )  z 1 , z 2  then z MOM 
z 2
 Mamdani Fuzzy models (cont.) 10
• Smallest of maximum zSOM
Amongst all z that belong to [z1, z2], the
smallest is called zSOM
• Largest of maximum zLOM
Amongst all z that belong to [z1, z2], the
largest value is called zLOM

Various defuzzification schemes for obtaining a crisp output

 Mamdani Fuzzy models (cont.) 11
– Example #1
Single input single output Mamdani fuzzy
model with 3 rules:
If X is small then Y is small  R1
If X is medium then Y is medium  R2
Is X is large then Y is large  R3
X = input [-10, 10] Y = output [0,10]
Using max-min composition (R1 o R2 o R3)
and centroid defuzzification, we obtain
the following overall input-output
curve
 12

Single input single output antecedent & consequent MFs

Overall input-output curve

– Example #2 (Mamdani Fuzzy models ) 13
Two input single-output Mamdani fuzzy
model with 4 rules:
If X is small & Y is small then Z is
negative large
If X is small & Y is large then Z is
negative small
If X is large & Y is small then Z is
positive small
If X is large & Y is large then Z is
positive large
 X = [-5, 5]; Y = [-5, 5]; Z = [-5, 5] with max-min
composition & centroid defuzzification, we can
14
determine the overall input output surface

Two-input single output antecedent & consequent MFs

 15

Overall input-output surface

 Mamdani Fuzzy models (4.2) (cont.) 16
• w1 = degree of compatibility between A & A’
• w2 = degree of compatibility between B & B’
• w1  w2 = degree of fulfillment of the fuzzy
rule (antecedent part) = firing strength
• Qualified (induced) consequent MFs
represent how the firing strength gets
propagated & used in a fuzzy implication
statement
• Overall output MF aggregate all the qualified
consequent MFs to obtain an overall output
MF
 Mamdani Fuzzy models (4.2) (cont.) 17

– One might use product for firing strength

computation

– One might use min for qualified consequent

MFs computation

– One might use max for MFs aggregation into

an overall output MF
 Conclusion: 18
To completely specify the operation of a Mamdani fuzzy
inference system, we need to assign a function for each
of the following operators:
• AND operator (usually T-norm) for the rule firing
strength computation with AND’ed antecedents
• OR operator (usually T-conorm) for calculating the
firing strength of a rule with OR’ed antecedents
• Implication operator (usually T-norm) for
calculating qualified consequent MFs based on
given firing strength
• Aggregate operator (usually T-conorm) for
aggregating qualified consequent MFs to generate
an overall output MF  composition of facts &
rules to derive a consequent
• Defuzzification operator for transforming an output
MF to a crisp single output value
Example: 19
 “product”  “sum”
Aggregate

This sum-product composition provides the following

theorem:
Final crisp output when using centroid defuzzification =
weighted average of centroids of consequent MFs
where:
w (rulei) = (firing strength)i * Area (consequent MFs)
Proof: Use the following:
 C' (z )  w1 C1 (z )  w 2 C2 (z )
and compute: zCOA (centroid defuzzification)
Conclusion: Final crisp output can be computed if:
– Area of each consequent MF is known
– Centroid of each consequent Mf is known
 Sugeno Fuzzy Models (4.3) 20
[Takagi, Sugeno & Kang, 1985]

 Goal: Generation of fuzzy rules from a given

input-output data set
 A TSK fuzzy rule is of the form:
“If x is A & y is B then z = f(x, y)”
Where A & B are fuzzy sets in the
antecedent, while z = f(x, y) is a crisp
function in the consequent
 f(.,.) is very often a polynomial function w.r.t.
x&y
 Sugeno Fuzzy Models (4.3) (cont.) 21

 If f(.,.) is a first order polynomial, then the

resulting fuzzy inference is called a first order
Sugeno fuzzy model

 If f(.,.) is a constant then it is a zero-order

Sugeno fuzzy model (special case of Mamdani
model)

 Case of two rules with a first-order Sugeno fuzzy

model
– Each rule has a crisp output
– Overall output is obtained via weighted average
– No defuzzyfication required
 22

The Sugeno fuzzy model

 Sugeno Fuzzy Models (4.3) (cont.) 23

Example 1: Single output-input Sugeno

fuzzy model with three rules

If X is small then Y = 0.1X + 6.4

If X is medium then Y = -0.5X + 4
If X is large then Y = X – 2

If “small”, “medium” & “large” are nonfuzzy

sets then the overall input-output curve is
a piece wise linear
24
However, if we have smooth membership 25
functions (fuzzy rules) the overall input-
output curve becomes a smoother one
 Example 2: 26
Two-input single output fuzzy model with 4 rules

R1: if X is small & Y is small then z = -x +y +1

R2: if X is small & Y is large then z = -y +3
R3: if X is large & Y is small then z = -x +3
R4: if X is large & Y is large then z = x + y + 2

R1  (x  s) & (y  s)  w1
R2  (x  s) & (y  l)  w2
R3  (x  l) & (y  s)  w3
R4  (x  l) & (y  l)  w4
Aggregated consequent  F[(w1, z1); (w2, z2); (w3, z3);(w4, z4)]

= weighted average
 27

Overall input-output surface

 28

The Tsukamoto fuzzy model

 Tsukamoto Fuzzy models (4.4) (cont.) 29

Example:
single-input Tsukamoto fuzzy model
with 3 rules
if X is small then Y is C1

if X is medium then Y is C2

if X is large then Y is C3
30
 Other Considerations (4.5) (cont.)
– Grid partition 31
Each region is included in a square area 
hypercube
Difficult to partition the input using the Grid in
the case of a large number of inputs. If we
have k inputs & m MFs for each  mk rules!!
– Tree partition
Each region can be uniquely specified along a
corresponding decision tree. No exponential
increase in the number of rules
– Scatter partition
Each region is determined by covering a
subset of the whole input space that
characterizes a region of possible occurrence
of the input vectors
 32

Various methods for partitioning the input space:

(a) Grid partition;
(b)tree partition;
(c) scatter partition.
 Fuzzy Modeling 33

We have covered several types of

fuzzy inference systems (FIS’s)

– A design of a fuzzy inference

system is based on the past known
behavior of a target system

– A developed FIS should reproduce

the behavior of the target system
 34
Fuzzy Modeling (cont.)
– Examples of FIS’s

• Replace the human operator that

regulates & controls a chemical
reaction, a FIS is a fuzzy logic
controller

• Target system is a medical doctor; a

FIS becomes a fuzzy expert system
for medical diagnosis
 35
– How to construct a FIS for a specific
application?

• Incorporate human expertise about the

target system: it is called the domain
knowledge (linguistic data!)

• Use conventional system identification

techniques for fuzzy modeling when input-
output data of a target system are
available (numerical data)
 36
i. Determine the number of linguistic terms
associated with each input & output variables (for
a Sugeno model, determine the order of
consequent equations
Part A describes the behavior of the target
system by means of linguistic terms
– General guidelines about fuzzy modeling
A. Identification of the surface structure

i. Select relevant input-output variables

ii. Choose a specific type of FIS
 37
B. Identification of deep structure

i. Choose an appropriate family of

parameterized MF’s
ii. Interview human experts familiar with the
target systems to determine the parameters of
the MF’s used in the rule base
iii. Refine the parameters of the MF’s using
regression & optimization techniques (best
performance for a plant in control!)

(i) + (ii): assumes the availability of human

experts
(iii): assumes the availability of the desired input-
output data set